Review comment updates

Signed-off-by: Akshay Tondak <[email protected]>
Xilinx · Oct 16, 2024 · d7c8282 · d7c8282
1 parent 4f2f7e9
commit d7c8282
Show file tree

Hide file tree

Showing 11 changed files with 18 additions and 26 deletions.
diff --git a/src/runtime_src/core/tools/common/tests/TestAIEReconfigOverhead.cpp b/src/runtime_src/core/tools/common/tests/TestAIEReconfigOverhead.cpp
@@ -134,8 +134,6 @@ TestAIEReconfigOverhead::run(std::shared_ptr<xrt_core::device> dev)
     logger(ptree, "Details", boost::str(boost::format("No. of iterations: %f") % itr_count));
   }
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   auto start = std::chrono::high_resolution_clock::now();
   for (int i = 0 ;i < itr_count ; i++){
     try{

diff --git a/src/runtime_src/core/tools/common/tests/TestCmdChainLatency.cpp b/src/runtime_src/core/tools/common/tests/TestCmdChainLatency.cpp
@@ -5,7 +5,6 @@
 // Local - Include Files
 #include "TestCmdChainLatency.h"
 #include "tools/common/XBUtilities.h"
-#include "TestValidateUtilities.h"
 #include "xrt/xrt_bo.h"
 #include "xrt/xrt_device.h"
 #include "xrt/xrt_hw_context.h"
@@ -138,8 +137,6 @@ TestCmdChainLatency::run(std::shared_ptr<xrt_core::device> dev)
   for (auto& run : runs)
     runlist.add(run);
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   auto start = std::chrono::high_resolution_clock::now();
   for (int i = 0; i < itr_count; ++i) {
     try {

diff --git a/src/runtime_src/core/tools/common/tests/TestCmdChainThroughput.cpp b/src/runtime_src/core/tools/common/tests/TestCmdChainThroughput.cpp
@@ -5,7 +5,6 @@
 // Local - Include Files
 #include "TestCmdChainThroughput.h"
 #include "tools/common/XBUtilities.h"
-#include "TestValidateUtilities.h"
 #include "xrt/xrt_bo.h"
 #include "xrt/xrt_device.h"
 #include "xrt/xrt_hw_context.h"
@@ -137,8 +136,6 @@ TestCmdChainThroughput::run(std::shared_ptr<xrt_core::device> dev)
   for (auto& run : runs)
     runlist.add(run);
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   auto start = std::chrono::high_resolution_clock::now();
   for (int i = 0; i < itr_count; ++i) {
     try {

diff --git a/src/runtime_src/core/tools/common/tests/TestDF_bandwidth.cpp b/src/runtime_src/core/tools/common/tests/TestDF_bandwidth.cpp
@@ -128,8 +128,6 @@ TestDF_bandwidth::run(std::shared_ptr<xrt_core::device> dev)
     logger(ptree, "Details", boost::str(boost::format("No. of iterations: %f") % itr_count));
   }
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   auto start = std::chrono::high_resolution_clock::now();
   for (int i = 0; i < itr_count; i++) {
     try {

diff --git a/src/runtime_src/core/tools/common/tests/TestGemm.cpp b/src/runtime_src/core/tools/common/tests/TestGemm.cpp
@@ -119,9 +119,9 @@ TestGemm::run(std::shared_ptr<xrt_core::device> dev)
   // Create 128KB Debug BO to capture TOPS data
   xrt::bo bo_result = xrt_core::bo_int::create_debug_bo(hwctx, 0x20000);
 
-  // wait until clock reaches the max frequency
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
+  // wait until clock reaches the max frequency. The performance metrics for a test
+  // are valid only when the clock reaches the max frequency.
+  uint64_t ipu_hclock = XBValidateUtils::wait_for_max_clock(dev);
 
   try {
     //run kernel

diff --git a/src/runtime_src/core/tools/common/tests/TestSpatialSharingOvd.cpp b/src/runtime_src/core/tools/common/tests/TestSpatialSharingOvd.cpp
@@ -118,8 +118,6 @@ boost::property_tree::ptree TestSpatialSharingOvd::run(std::shared_ptr<xrt_core:
     }
   }
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   // Measure the latency for running the test cases in parallel
   auto start = std::chrono::high_resolution_clock::now(); 
 

diff --git a/src/runtime_src/core/tools/common/tests/TestTCTAllColumn.cpp b/src/runtime_src/core/tools/common/tests/TestTCTAllColumn.cpp
@@ -125,8 +125,6 @@ TestTCTAllColumn::run(std::shared_ptr<xrt_core::device> dev)
     logger(ptree, "Details", boost::str(boost::format("No. of iterations: %f") % itr_count));
   }
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   auto start = std::chrono::high_resolution_clock::now();
   try {
     auto run = kernel(host_app, bo_ifm, NULL, bo_ofm, NULL, bo_instr, instr_size, NULL);

diff --git a/src/runtime_src/core/tools/common/tests/TestTCTOneColumn.cpp b/src/runtime_src/core/tools/common/tests/TestTCTOneColumn.cpp
@@ -124,8 +124,6 @@ TestTCTOneColumn::run(std::shared_ptr<xrt_core::device> dev)
     logger(ptree, "Details", boost::str(boost::format("No. of iterations: %f") % itr_count));
   }
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   auto start = std::chrono::high_resolution_clock::now();
   try {
     auto run = kernel(host_app, bo_ifm, NULL, bo_ofm, NULL, bo_instr, instr_size, NULL);

diff --git a/src/runtime_src/core/tools/common/tests/TestTemporalSharingOvd.cpp b/src/runtime_src/core/tools/common/tests/TestTemporalSharingOvd.cpp
@@ -90,8 +90,6 @@ TestTemporalSharingOvd::run(std::shared_ptr<xrt_core::device> dev) {
     }
   };
 
-  int ipu_hclock = 0;
-  XBValidateUtils::wait_for_max_clock(ipu_hclock, dev);
   // Measure the latency for running the test cases in parallel
   auto start = std::chrono::high_resolution_clock::now(); 
   // Create two threads to run the test cases

diff --git a/src/runtime_src/core/tools/common/tests/TestValidateUtilities.cpp b/src/runtime_src/core/tools/common/tests/TestValidateUtilities.cpp
@@ -135,12 +135,22 @@ get_instr_size(const std::string& dpu_file) {
   return size;
 }
 
-void
-wait_for_max_clock(int& ipu_hclock, std::shared_ptr<xrt_core::device> dev) {
+/**
+ * @brief Waits for the IPU clock frequency to reach the target maximum clock frequency.
+ *
+ * This function queries the device for the target maximum clock frequency and then
+ * continuously checks the current IPU clock frequency until it reaches the target.
+ *
+ * @param dev A shared pointer to the xrt_core::device.
+ * @return The IPU clock frequency when it reaches the target maximum clock frequency.
+ */
+uint64_t
+wait_for_max_clock(std::shared_ptr<xrt_core::device> dev) {
   uint64_t target_h_clock_freq = 0;
+  uint64_t ipu_hclock = 0;
   auto res_info = xrt_core::device_query<xrt_core::query::xrt_resource_raw>(dev);
   if (res_info.empty())
-    return;
+    return ipu_hclock;
 
   for (auto &res : res_info)
   {
@@ -158,6 +168,6 @@ wait_for_max_clock(int& ipu_hclock, std::shared_ptr<xrt_core::device> dev) {
     }
     std::this_thread::sleep_for(std::chrono::milliseconds(50));
   }
+  return ipu_hclock;
 }
-
 }// end of namespace XBValidateUtils
diff --git a/src/runtime_src/core/tools/common/tests/TestValidateUtilities.h b/src/runtime_src/core/tools/common/tests/TestValidateUtilities.h
@@ -66,7 +66,7 @@ namespace XBValidateUtils{
 
 void init_instr_buf(xrt::bo &bo_instr, const std::string& dpu_file);
 size_t get_instr_size(const std::string& dpu_file);
-void wait_for_max_clock(int&, std::shared_ptr<xrt_core::device>);
+uint64_t wait_for_max_clock(std::shared_ptr<xrt_core::device>);
 
 } //End of namespace XBValidateUtils
 #endif